Ask your assistant, not a guess

One URL, no install. It works without a key for 20 calls a day; add a free key as Authorization: Bearer for 200 a day, or a paid plan for more.

https://llmbottleneck.com/mcp

Claude

Settings → Connectors → Add custom connector → paste the URL. In Claude Code:

claude mcp add --transport http llmbottleneck https://llmbottleneck.com/mcp

ChatGPT

Settings → Apps & Connectors → Advanced → Developer mode → Create → paste the URL, no authentication.

Cursor, VS Code, Windsurf and other clients

{
  "mcpServers": {
    "llmbottleneck": { "url": "https://llmbottleneck.com/mcp" }
  }
}

A client that only runs local commands can use the stdio bridge — the same file as the terminal tool below, at the path you saved it:

{
  "mcpServers": {
    "llmbottleneck": { "command": "node", "args": ["/path/to/llmbottleneck.mjs", "mcp"] }
  }
}

Tools

can_i_run — Can this device run this model?

Whether an open-weight LLM fits in a GPU's or Mac's memory, at every quantization or one you name, with the memory needed and an estimated decode speed. Sizes come from published files and manufacturer specifications.

what_runs_on — What runs on this device?

The open-weight models a GPU or Mac holds entirely in its memory, with the format, memory needed and estimated speed.

hardware_for_model — What hardware runs this model?

The single GPUs and Macs that hold a model at a quantization (default Q4_K_M), smallest memory first, optionally only those reaching a decode speed.

cheapest_way_to_run — Cheapest way to run this model

The cheapest rented GPU machine that holds the model (Vast.ai, RunPod prices) and the per-token API price, with a link to compare owning, renting and per-token for your own usage.

search_catalogue — Search the catalogue

Find the exact ids of catalogued models and devices from a partial name.

Every tool is read-only and answers from the pinned catalogue. Each answer ends with the page it came from, so what the assistant tells you can be checked.

Node 18 or newer. One file to download, no dependencies and nothing to install:

curl -fsSL https://llmbottleneck.com/v1/clients/llmbottleneck.mjs -o llmbottleneck.mjs

node llmbottleneck.mjs                         # detect your GPU, list what it runs
node llmbottleneck.mjs can-i-run "Qwen3-32B"   # every format on the detected GPU
node llmbottleneck.mjs hardware "Llama 3.3 70B" --min-tps 20
node llmbottleneck.mjs cheapest "Qwen3-32B"

Help calibrate the speeds. With llama.cpp installed, node llmbottleneck.mjs bench --model file.gguf measures your machine and shows the estimate beside it; --submit sends the numbers for review. Nothing is published without review, and only the GPU name, the file name and size and the benchmark’s figures are sent.